In electric power systems, generators can be modeled by two components: a mechanical part (turbine) and an electric part (voltage source). The mechanical power input and electric power output are identical in steady-state. Whenever a contingency occurs, imbalances between these quantities force some generators to speed up while others slow down. Generators may lose synchronization and, consequently, the system collapses. The core of the problem is the fact that mechanical adjustments are much slower than electric changes.
A new family of solid-state devices, called Flexible AC Transmission Systems (FACTS), allows fast and direct control on some parameters of power systems, such as line and shunt impedances. FACTS devices can indirectly affect the electric part of generators and quickly respond to imbalances.
We are proposing to devise distributed control strategies for FACTS devices. We conjecture that geographically distributed controllers using only locally available state information can achieve performance comparable to centralized controllers. Strategies may be learned by trial-and-error or from training data generated by "an ideal" centralized controller. Autonomous agents follow these strategies and cooperate with their neighbors to avoid conflicting actions. These agents may cooperate by estimating how their actions affect their neighbors, communicating with them, and taking combined actions.
This research is in its nascent stages and only preliminary results have been achieved. Distributed controllers were synthesized for a few power systems by learning from training data generated by a centralized controller. A collaboration heuristic also illustrates how distributed controllers can take combined actions.
My thesis proposal and the transparencies of my presentation can be downloaded by clicking on the items below: